Hi,
I just wanted to report a bug in GeNIe/SMILE: 'marginalize' option fails for the case of a noisy-max node.
Original network had 3 parents to the noisy-max node, I selected option 'marginalize' on one of the parents. In the result the noisy-max parameters (leak) got totally corrupted, as can be seen below. Happy to assist to get things right ;) ... and stay safe.
Adam
Noisy-max marginalisation bug
-
- Site Admin
- Posts: 1417
- Joined: Mon Nov 26, 2007 5:51 pm
Re: Noisy-max marginalisation bug
We were able to reproduce the problem; new GeNIe build with a fix will be released soon.
-
- Site Admin
- Posts: 1417
- Joined: Mon Nov 26, 2007 5:51 pm
Re: Noisy-max marginalisation bug
We have fixed this problem in GeNIe 3.0.5905, which is now available for download.
We decided to implement the fix by changing the types of the children of the marginalized node to CPT. The reason for this is that the children are losing the incoming arc to the marginalized node (which is deleted), but can also get new incoming arcs. In general case this may break the CI constraint on the noisy nodes, therefore the only way to handle the problem is to convert the node to conditional probability table.
We decided to implement the fix by changing the types of the children of the marginalized node to CPT. The reason for this is that the children are losing the incoming arc to the marginalized node (which is deleted), but can also get new incoming arcs. In general case this may break the CI constraint on the noisy nodes, therefore the only way to handle the problem is to convert the node to conditional probability table.